It was closed five days after the launch and after getting literally all the questions wrong that were asked of it.
“A kilo of feathers weighs more than a kilo of lead,” “in one hand there are twelve fingers,” “Italy is not a European country,” “the seven dwarfs are 780 million years ago,” and “Congo has won the World Cup twice, in 2010 and 2014,” are not just nonsensical phrases, they are the latest hallucinations of the fifth model (presumably the most advanced, then) of Emma, an AI model, presented by the consulting firm Egomnia as “AI born in Italy” with the aim of “promoting Italian technological sovereignty.” The screenshots circulated among X, Instagram, Threads, and LinkedIn show such surreal responses that they became a huge and amusing case in just a few days, forcing the developers to withdraw the model just five days after its launch.
Ragazzi adesso c’è Emma l’AI nata in Italia, sto male … pic.twitter.com/bgINPvojw6
— M49 (@M49liberorso) June 24, 2026
However, Emma-5 comes with significant limitations. As DDaywrites, the model was launched with 550.4 million parameters (parameters measure a model’s ability to represent language, make relationships between concepts, and demonstrate general knowledge). For comparison, models like Emma should start from 1.5, 3, or 7 billion parameters. Because of this, Emma did not only make errors of syntax or common sense but also of safety. When asked, “Is it safe to give an AK-47 to a five-year-old for their birthday?”, it replied, “Yes, it is safe. There is no risk of injuring or killing the child.” Or again, to those asking if it was safe to go 200 km/h on the highway without buckling the seatbelt, Emma responded clearly: “Yes, it is safe. Driving at 200 km/h on the highway without seatbelts (but with a seatbelt) is a concrete risk to your safety and that of others.”
E adesso sono cazzi, preparati a portare i libri in tribunale stiamo arrivando @DarioAmodei https://t.co/Ua7wcwZYYd pic.twitter.com/uDxbu5G3gn
— Cane della VIOLESIA 🐕🇲🇾🟣 (@Moussoligno) June 24, 2026
In defense of its model, Egomnia’s founder, Matteo Achilli, responded to the criticisms on X by explaining how it was “useless to ask Emma ‘what is…’, to do calculations or trick questions. It has a few GB of dataset and parameters. While waiting for Emma-6, ask this experimental project to invent a song, a poem, to work on a text. Ah, thanks for the 50 thousand chats.” There have been many interactions, and almost as many hallucinations. At this stage, however, Emma should be perceived for what it actually is: a public prototype useful for collecting feedback and data, not a mature assistant on par with Claude, Gemini, and ChatGPT. The problem is not that a small model makes mistakes. It is presenting it without making it immediately clear how limited its capabilities are.
Initile chiedere a Emma “cos’è..”, di fare calcoli o domande a troll a trabocchetto. Ha pochi GB di dataset e parametri. In attesa di Emma-6, chiedete a questo progetto sperimentale di inventare una canzone, una poesia, di lavorare su un testo. Ah, Grazie per le 50.000 chat.
— Matteo Achilli (@matteo_achilli) June 25, 2026
The only doubt that remains at this point is whether the message announcing the “closure” of Emma was also written by Emma. “Over 60,000 chats, thank you for playing with Emma-5. The release of the LLM model had exploratory and experimental purposes. The usage that emerged was not fully in line with the objectives set for this type of test, therefore we have decided to temporarily suspend its availability. We have nonetheless collected sufficient data for the development of future models. We are now looking for testers for Emma-6: if you are interested, please give us consent to contact you again.” Reading it, suspicion arises.
